Tarte

A secondary layer for pie for disambiguation

What it aims to do

• This tagger is supposed to come as a secondary layer for lemma that should be disambiguated.
• Its core object (Tarte) should filter things that need to be disambiguated
• Its training capacities should reorganize a training set so that it dispatch training samples across all sets and it should not care about sample not containing unambiguous tokens.
• It takes POS, lemma context and form characters into the network to predict the disambiguated form.

Notes

• Given that not all sentences will have things to disambiguate, pretraining vector might be an important task. It is possible with PyTorch to load Gensim data easily. This would require to generate temps file where lemma AND pos are fed to fake sentences.